[LLVMdev] Strange ppc failure.

[Hal Finkel]

----- Original Message -----
> From: "Richard Pennington" <rich at pennware.com>
> To: "LLVM Developers Mailing List" <llvmdev at cs.uiuc.edu>
> Sent: Saturday, May 24, 2014 10:01:22 PM
> Subject: [LLVMdev] Strange ppc failure.
> 
> As part of my testing for my ELLCC cross compiler tool project
> (http://ellcc.org), I use clang/LLVM to compile itself for various
> targets
> (arm, armeb, i386, mips, mipsel, ppc, and x86_64). Today, after a TOT
> update the ppc cross build failed strangely. I started to get
> assembly errors:
> 
> [~/ellcc/llvm/lib/Target/PowerPC] dev% ~/ellcc/bin/ppc-elf-as
> PPCInstrInfo.s
> PPCInstrInfo.s: Assembler messages:
> PPCInstrInfo.s:1795: Error: operand out of range (49 is not between 0
> and 31)
> PPCInstrInfo.s:17893: Error: operand out of range (73 is not between
> 0
> and 31)
> PPCInstrInfo.s:17953: Error: operand out of range (49 is not between
> 0
> and 31)
> 
> Here is the assembly starting at 17980 (other cases are similar):
> .LBB133_139:
>          lbz 3, 75(22)
>          lwz 0, 140(1)
>          add 18, 22, 73                  <-- error line
>          rlwinm 0, 0, 20, 0, 31
>          andi. 3, 3, 1
>          mtcrf 16, 0
>          bc 4, 1, .LBB133_142
>          lwz 3, 164(1)
>          lwz 6, 236(1)
>          mr 4, 16
>          li 5, 374
>          bl __assert_fail
>          b .LBB133_142
> .LBB133_141:
>          lwz 0, 140(1)
>          addi 18, 22, 73                 <-- odd similarity
> 
> It looks (to me, brain hazy with a bad head code) that the add in the
> error
> case might want to be an addi.

Indeed. Can you grab the input LLVM IR and file a bug report, and we'll see if we can reproduce from that (pass -S -emit-llvm)? If we can reproduce this using the IR and llc, then it should be possible to narrow down the problem using bugpoint.

Thanks,
Hal

> 
> My previous TOT revision that worked for sure was 208501. My current
> test version is 209581.
> 
> Any thoughts? Even suggestions about how to submit bug report (simple
> test case) would be welcome.
> 
> -Rich
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
> 

-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory